Alhazen-OCR

Alhazen-OCR is an Arabic-first OCR vision-language model: a QLoRA adapter on top of Qwen3-VL-2B-Instruct, fine-tuned on context212/atlas-ocr-data — a license-clean mix of Arabic synthetic documents, invoices, and handwritten text (KHATT).

Named after Ibn al-Haytham (Alhazen), the 11th-century pioneer of optics.

Benchmark

Evaluated with CER/WER (lower is better) on the held-out eval split of the training data and on the external ahmedheakl/arocrbench_khatt benchmark (100 samples each, greedy decoding):

Model	Eval CER ↓	Eval WER ↓	KHATT-bench CER ↓	KHATT-bench WER ↓
Qwen3-VL-2B-Instruct (base)	0.653	0.951	1.498	1.594
Alhazen-OCR (1 epoch)	0.368	0.567	2.761	1.707

Reading the numbers: fine-tuning cuts character errors nearly in half on in-distribution documents (0.65 → 0.37 CER) and word errors from 0.95 → 0.57. On the external KHATT handwriting benchmark the model still struggles — this is a first smoke run on 20k samples; scaling data and training is expected to close the gap. Published as a work-in-progress baseline.

Usage

import torch
from unsloth import FastVisionModel

model, tokenizer = FastVisionModel.from_pretrained(
    "context212/alhazen-ocr",
    load_in_4bit=True,
)
FastVisionModel.for_inference(model)

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": image},  # a PIL.Image
        {"type": "text", "text": "Extract all the text from this image, preserving the original reading order."},
    ],
}]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=512, temperature=0.0, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training

• Base: unsloth/Qwen3-VL-2B-Instruct (4-bit)
• Method: QLoRA via unsloth — r=16, rsLoRA, vision + language layers (17.4M trainable params, 0.81%)
• Data: 20,000 samples from context212/atlas-ocr-data (60% synthetic documents, 20% invoices, 20% handwriting)
• Schedule: 1 epoch, lr 2e-5, cosine, bf16, batch 16, seq 2048 — single A100 80GB, ~35 min
• Checkpoint selection: per-epoch eval on held-out split + external benchmark; epoch 1 won (later epochs diverge — see the repo for details)

Data sources & licenses

Source	License
loay/arabic-ocr-synthetic-scans-faker-300k	CC-BY-4.0
KhalfounMehdi/arabic-latin-invoices-synthetic	CC-BY-4.0
johnlockejrr/KHATT_v1.0_dataset	MIT
sherif1313/Historical-Arabic-Handwritten-OCR	Apache-2.0

Eval benchmark: ahmedheakl/arocrbench_khatt (MIT).

Limitations

• First smoke run (20k samples, 1 epoch) — handwriting and out-of-distribution documents remain weak (see KHATT-bench column)
• Trained on line/paragraph-level crops; full-page layout parsing untested
• Arabic-first; other scripts not evaluated

Code

Training, data pipeline and evaluation: github.com/context212/atlas-ocr